What data for this category looks like
Data for this category comes from product digital models and GD&T specifications, process route documents, coordinate measuring machine data, on-site trial feedback, and metrology system records. Daily generated documents include tooling requirement documents, positioning datum schemes, fixture engineering drawings, gauge operation instructions, acceptance reports, calibration records, and change histories. Data updates are carried out on a per-project basis, triggered by product changes, and also completed through periodic calibration updates. Most documents contain geometric dimension annotations, tolerance parameters, operation steps, and verification records. Fields are often linked to specific workpiece models and tooling numbers.
What constraints these characteristics impose on knowledge base retrieval and recall
Product digital models and GD&T data include high-precision geometric parameters. Require retrieved content to match exact datum and tolerance fields. Avoid fuzzy matching to prevent audit deviations. Store multiple document types together. Differentiate retrieval weights for fixture drawings, calibration records, and other document types. Prioritize core audit basis for recall. Updates triggered by projects and changes require retrieval results to link to the latest change history data. Avoid using expired calibration or acceptance records. On-site trial feedback and metrology data have timeliness requirements. Prioritize recently updated verification documents for recall. This ensures the accuracy of audit results. A single retrieval may need to cover multiple related documents. Balance recall scope and content redundancy. Avoid overly long context to prevent reduced processing efficiency.
Configuration setup
This configuration is based on FastGPT v4.17.0. Specific parameter settings are as follows:
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.65–0.75 | Matches high-precision tooling datum and tolerance parameters. A value too low allows irrelevant content to be included. A value too high leads to insufficient recall results |
limit | Top 8–12 results | Covers related documents such as fixture drawings and calibration records under a single project, while controlling context length |
searchMode | mixedRecall | Balances exact matching of GD&T parameters and semantic association of document types |
usingReRank | Enabled | Re-sorts initial recall results by relevance to the current audit task. Prioritizes latest calibration records and change histories |
datasetSearchUsingExtensionQuery | Enabled | Expands retrieval to linked change histories and acceptance reports based on tooling numbers or workpiece models entered by users |
Three common mistakes
- Phenomenon: After adjusting
similarityto the minimum andlimitto the maximum, the number of recall results remains fixed with no growth. Cause: Some datasets have a global upper limit set forlimit, or the retrieval scope is only bound to specifiedcollectionIdsand does not cover all related documents. - Phenomenon: Fixture engineering drawing tables in the knowledge base lose their table formatting in the output content after retrieval and recall. Cause: The document format retention configuration is not enabled, or the table structure was not parsed correctly during upload.
- Phenomenon: It is not possible to call a dedicated knowledge base for different business scenarios. For example, it is not possible to separately call the calibration record knowledge base for compliance verification. Cause: The
collectionIdsparameter is not configured to bind the corresponding knowledge base collection, or retrieval rules are not used to distinguish the call scope of different datasets.
How to confirm configuration is correct
- Enter an audit query that includes a specific tooling number and tolerance parameters. Check whether the retrieved results include the fixture drawings and calibration records for the corresponding model.
- Adjust the
similarityparameter to the 0.6 and 0.7 ranges. Observe changes in the number and relevance of recall results to confirm the configuration is effective. - Upload a small tooling requirement document and a large change history. Verify whether the retrieval task can process documents sorted by size, or confirm the configuration status of the upload pause function.
- Check whether the table formatting of the original document is retained in the output content. Confirm that the format parsing configuration is correct.
The values provided are common starting points and should be measured against the reader's own samples.